The dynamical memory of tidal stellar streams: Joint inference of the Galactic potential and the progenitor of GD-1 with flow matching

arXiv:2512.04600 · astro-ph.GA, physics.class-ph, physics.data-an, physics.space-ph · Submitted 2025-12-04 · Read on arXiv

Listen

Radio episode about this paper

Transcript

Introduction to the show: ident: Astrophysics Radio. Generated commentary on the latest astrophysics papers.

Vera: We're looking at the paper: "The dynamical memory of tidal stellar streams".

Jocelyn: Here's the gist of the paper. The gist Stellar streams offer one of the most sensitive probes of the Milky Way’s gravitational potential, as their phase-space morphology encodes both the tidal field of the host galaxy and the internal structure of their progenitors.

Vera: Framework and Goal The aim is to move beyond traditional techniques by constructing a fully Bayesian, likelihood-free posterior over both host-galaxy parameters and progenitor properties, thereby capturing the intrinsic coupling between tidal stripping dynamics and the underlying potential.

Jocelyn: To achieve this, researchers leverage Flow Matching and Simulation-Based Inference (SBI) to jointly infer the parameters of the GD-1 progenitor and the global properties of the Milky Way potential.

Vera: This approach allows for joint inference on progenitor and Galactic parameters, capturing complex dependencies that are difficult to model with classical likelihood-based methods.

Title and authors: Vera: Basically, they are using Flow Matching and Simulation-Based Inference to figure out the parameters for both the host galaxy potential and the progenitor object simultaneously.

Jocelyn: They are going beyond traditional methods that try to fit everything one piece at a time, which is what we usually see in orbit-fitting approaches.

Subrahmanyan: The paper focuses on recovering physical parameters theta that govern how a globular cluster evolves within a Milky Way-like gravitational field (<ref:2512.04600#pg2>).

Vera: They model the host using an NFW halo, a Miyamoto-Nagai disc, and an exponential bulge, and they use a Plummer sphere for the progenitor model (<ref:2512.04600#pg3>).

Jocelyn: The key innovation is training a neural density estimator to map the observed stream data directly to these combined parameters.

Subrahmanyan: They achieve this by learning a vector field, v phi, that describes how the distribution of particles evolves over time in that potential (<ref:2512.04600#pg3>).

Vera: The math behind it uses conditional Flow Matching loss, LCFM, which is designed to learn this transport from a base Gaussian distribution to the actual posterior distribution (<ref:2512.04600#pg3>).

Jocelyn: It’s an amortized inference pipeline, meaning once you train the network, it can give you inferences about new data quickly without needing lengthy sequential training processes.

Subrahmanyan: This is significant because it allows for joint constraints on progenitor and Galactic parameters, capturing the coupling between tidal stripping dynamics and the host potential (<ref:2512.04600#pg1>).

Vera: So, in simple terms, they’re using machine learning to jointly map a complex physical process—tidal disruption—to the underlying gravitational landscape of a galaxy and its member cluster.

The paper's summary: Jocelyn: One major improvement is achieving an amortized inference pipeline for Galactic structure parameters and progenitor properties from single stellar stream phase-space data without requiring sequential training.

Vera: That means you get a significant computational gain because you don't have to run a lengthy training process every time you want to infer something new.

Subrahmanyan: And the second major point is the joint constraint on progenitor physics, where they capture that intrinsic coupling between tidal stripping dynamics and the underlying potential (<ref:2512.04600#pg1>).

Jocelyn: This moves beyond just fitting orbits; it lets them constrain parameters that govern both the host galaxy’s gravity and the internal structure of the progenitor, which is a big step forward in understanding tidal interactions.

Vera: They also use rigorous validation through simulation-based checks, including Predicted Percentile plots for calibration and Tests of Accuracy with Random Point to check coverage (<ref:2512.04600#pg2>).

Subrahmanyan: And they use Posterior Predictive Checks to ensure the inferred parameters are self-consistent when they forward model the results against mock observations (<ref:2512.04600#pg3>). This level of validation is important for trusting the output.

Jocelyn: So, in short, it’s about making a scalable pathway for dynamical inference that is flexible and avoids the computational expense of gold standard MCMC methods when dealing with these high-dimensional problems.

The paper's improvements: Vera: The main implication is that this approach allows us to jointly constrain parameters for both the host galaxy potential and the progenitor object, using just one stream.

Jocelyn: It shows that these streams carry detailed information about both the environment they are in and their origin, which is much richer than what we could get by studying either component separately.

Subrahmanyan: The ability to capture those complex correlations between parameters, such as between host potential parameters and progenitor mass scale length parameters, reflects physical expectations derived from simulations (<ref:2512.04600#pg2>).

Vera: So, the technique provides a controlled way to quantify the information that a single cold stellar stream carries about both the progenitor properties and the host galaxy potential in a simulation setting.

Jocelyn: This is an efficient way to get these constraints, and it sets a new standard for how we can use dynamical data from streams.

Subrahmanyan: It’s a powerful tool for testing structure formation models because it lets us probe the physical reality of tidal stripping in action (<ref:2512.04600#pg1>).

Vera: That's all we have on this paper today, but keep an eye on how these techniques apply to other streams.

Conclusion: Vera: So we’ve been talking about how they used Flow Matching to jointly figure out both the host galaxy potential and the progenitor of GD-one using just one stellar stream, in this paper called "The dynamical memory of tidal stellar streams: Joint inference of the Galactic potential and the progenitor of GD-one with flow matching."

Jocelyn: Right. It’s about using that phase-space data to constrain both where you are and where it came from simultaneously.

Subrahmanyan: The core idea is capturing that intrinsic coupling between the tidal stripping dynamics and the underlying potential, which is something traditional methods usually miss.

Vera: Exactly, so they model the Milky Way with an NFW halo, a Miyamoto-Nagai disc, and an exponential bulge for the host structure.

Jocelyn: And for the progenitor of GD-one they use a Plummer sphere model which is pretty standard for these kinds of disrupted globular clusters.

Subrahmanyan: The methodology involves training a Neural Density Estimator using conditional Flow Matching to learn a vector field that transports particles from some base distribution to the actual posterior distribution.

Vera: That’s the technical heavy lifting, so they use this AI-driven approach for amortized inference, meaning it doesn't need sequential training every time you want an answer.

Jocelyn: It’s fast and scalable, which is crucial when dealing with these high-dimensional problems involving many parameters for both the galaxy and the cluster.

Subrahmanyan: What this means cosmologically is that we can actually quantify how much information a single cold stellar stream carries about both the progenitor properties and the host potential.

Vera: That’s pretty powerful because it gives us a way to test structure formation models against real observational data from streams.

Jocelyn: And they showed strong agreement with mock observations using posterior predictive checks, which adds a good layer of confidence to their results on the model calibration.

Subrahmanyan: The paper also highlights some interesting parameter correlations, like the relationship between the NFW and Miyamoto-Nagai parameters for the host potential.

Vera: It shows that there's a weak positive correlation between those masses themselves, which is something you wouldn't expect right away.

Jocelyn: So it’s not just fitting things; it’s mapping out these physical dependencies across different components of the system.

Subrahmanyan: The limitation they point out is that the method relies on a set of predefined models for both the host and progenitor, so it's constrained by those assumptions.

Vera: That’s true. It doesn't explore completely novel physical scenarios outside those initial models, but it’s very strong within that framework.

Jocelyn: Still, the fact that they can constrain these parameters so well using just one stream is a big deal for observational astronomy.

Subrahmanyan: It suggests we need to look for these kinds of joint constraints in other types of streams or even other galaxy systems to confirm this kind of information transfer.

Vera: So, "The dynamical memory of tidal stellar streams: Joint inference of the Galactic potential and the progenitor of GD-one with flow matching" gives us a concrete way to extract more physics from these streams.

Jocelyn: It’s a solid paper for anyone looking to use simulation-based methods for complex dynamical problems.

Subrahmanyan: Next time, we’ll be talking about how polarization data from AGNs, like the VAPOLA survey, is evolving over time.

Giuseppe Viterbo, Tobias Buck

University Heidelberg

astro-ph.GA, physics.class-ph, physics.data-an, physics.space-ph

Submitted: 2025-12-04

Updated: 2025-12-04

Comments: submitted to A&A, comments welcome, all source code to reproduce this work can be found on GitHub under the url: https://github.com/vepe99/sbi-sim/tree/odisseo_branch . The simulator \textsc{\texttt{Odisseo}} is available on GitHub at the following url: https://github.com/vepe99/Odisseo

Journal ref: A&A 707, A363 (2026)

DOI: 10.1051/0004-6361/202558358

Code: https://github.com/vepe99/sbi-sim

License: http://creativecommons.org/licenses/by/4.0/

Importance score: 89/100

The gist: The gist Stellar streams offer one of the most sensitive probes of the Milky Way’s gravitational potential, as their phase-space morphology encodes both the tidal field of the host galaxy and the

Key concepts

Flow Matching: A technique used in AI to learn the path (vector field) that transforms an initial simple distribution (like a Gaussian) into a complex target distribution (the posterior). This allows the system to efficiently map observed stream data directly onto the probability of different host and progenitor parameters.
Simulation-Based Inference: A machine learning approach that uses simulations to train a model capable of inferring unknown parameters from observational data, rather than relying on traditional likelihood calculations. It allows for joint inference on multiple, complex variables simultaneously.
Flow Matching Loss (LCFM): The specific mathematical loss function used during training. It ensures the learned vector field accurately transports particles from a starting distribution to the true posterior distribution of parameters, guiding the neural network to learn the correct physical relationships.
SetTransformer Architecture: A neural network structure designed to handle high-dimensional data by using Multi-Layer Perceptrons (MLPs) for embedding and Self-Attention Blocks (SABs) to capture interactions among particle features, making it suitable for complex inference tasks.

Terminology

Flow Matching — 4× in this paper · explained in 20 episodes
Conditional Flow Matching — 1× in this paper · explained in 2 episodes
Degeneracies — 1× in this paper · explained in 2 episodes
Simulation-Based Inference (SBI) — 1× in this paper · explained in 4 episodes
Simulation-based Inference — 1× in this paper · explained in 3 episodes

Summary

The gist Stellar streams offer one of the most sensitive probes of the Milky Way’s gravitational potential, as their phase-space morphology encodes both the tidal field of the host galaxy and the internal structure of their progenitors.

Framework and Goal

The aim is to move beyond traditional techniques by constructing a fully Bayesian, likelihood-free posterior over both host-galaxy parameters and progenitor properties, thereby capturing the intrinsic coupling between tidal stripping dynamics and the underlying potential. To achieve this, researchers leverage Flow Matching and Simulation-Based Inference (SBI) to jointly infer the parameters of the GD-1 progenitor and the global properties of the Milky Way potential. This approach allows for joint inference on progenitor and Galactic parameters, capturing complex dependencies that are difficult to model with classical likelihood-based methods.

Host Galaxy Model

The gravitational potential of the Milky Way is modeled using a three-component representation consisting of a spherical dark-matter halo following a Navarro–Frenk–White profile (NFW), an axisymmetric disc described by a Miyamoto-Nagai (MN) model, and a spherical bulge with an exponential cut-off. This framework consists of:

A spherical dark-matter halo following a Navarro–Frenk–White profile (NFW) with density profile ρ(r) = ρ0 r s / (r s + r 2) 1, where r is the spherical radii coordinates, rs is the scale radius, and ρ0 is the central density.
An axisymmetric disc described by a Miyamoto-Nagai (MN) with potential Φ(R,z) = -GMMN / (R squared + (a + √z squared + b 2) 2), where R,z being the radii and height in cylindrical coordinates, MMN is the total mass of the disk, and a, b are respectively the scale length and height.
A spherical bulge with an exponential cut-off with a density profile ρ(r) = ρ0 / r 1 rc exp -(r/rc) squared, where rc and α being respectively the cut-off radius and the power-law exponent.

Progenitor Model and Simulation

The progenitor of GD-1 is widely attributed to be a disrupted globular cluster, which is well approximated as spherically symmetric, self-gravitating stellar ensembles, for which a Plummer sphere provides an analytically convenient and physically motivated model. The initial conditions are sampled by drawing star particles from a Plummer potential centered on the origin and then shifting them in phase space by (x c 0, v c 0). The simulation involves tracing the trajectory back to tend, as a single particle with mass MPlummer in the chosen gravitational potential to the initial phase space position (x c 0, v c 0).

Inference Method: Flow Matching

The inference is performed by training a Neural Density Estimator (NDE) of the posterior distribution p(θ dobs), where θ = (θprog, θhost) and dobs = (ϕ1, ϕ2,r, vϕ1 cos ϕ2, vϕ2, vr). The method uses conditional Flow Matching to learn the vector field that transports a base Gaussian distribution into the posterior p(θ d), enabling efficient, amortized inference directly from stream phase-space data. This is achieved by training a Neural Network (NN) to approximate the vector field vϕ, parametrized by a Neural Network with weights ϕ, that describes these trajectories by solving an ordinary differential equation (ODE). The loss function used is the sample-conditional flow matching loss (CFM) LCFM = Et∼U[0,1],d∼p(dθ),θt∼qt(θtθ1)∥vϕ(t, d, θt) − ut(θt θ1)∥2.

Architecture and Training

The architecture adopted is the SetTransformer architecture introduced by Lee et al. 2019, which extends this framework to infer the vector field vϕ, conditioned on the flow matching integration time t, the model parameters θt, and the particle observations d. Each input is first embedded into a latent representation via Multi-Layer Perceptrons (MLPs). The integration time t is further encoded through a sinusoidal time embedding following Ho et al. 2020. The particle features are processed by a stack of Self-Attention Blocks (SABs) with residual connections, which capture interactions among particles.

Validation and Results

The results are validated using several metrics, including Percentile–percentile (P–P) plots to evaluate calibration and Tests of Accuracy with Random Point (TARP) to study posterior coverage. The model is reported to be well-calibrated on almost all the marginals and the joint, suggesting good calibration for θhost. The results suggest that we can strongly constrain these parameters, since the true value always lies inside the 16-84 percentile. The posterior predictive check shows strong agreement with the mock observation of GD1.

Conclusion

The work presents the first use of Simulation Based Inference applied to the GD1 stream to jointly constrain parameters of the progenitor and the host potential. This approach enables an amortized inference that does not require sequential training. The authors capture strong correlations between masses and scale lengths parameters (MNFW − rNFW; MMN − aMN;rNFW − aMN), and surprisingly a weak positive correlation between the masses (MNFW − MMN). They leverage the freedom given by the Flow Matching technique to adopt a robust Transformer architecture, capable of handling this high dimensional inference task. The aim is to quantify, in a controlled simulation setting, the information that a single cold stellar stream (GD-1 analogues) carries about both the progenitor properties and the host galaxy potential.

Code Availability

The code is publicly released to reproduce all the figures via Github. The code is publicly released to reproduce all the figures via Github. The code is publicly released to reproduce all the figures via Github.

Acknowledgements

This project was made possible by funding from the CarlZeiss-Stiftung. This work was supported by the Deutsche Forschungsgemeinschaft (DFG, German Research Foundation) under Germany’s Excellence Strategy EXC 2181/1 - 390900948 (the Heidelberg STRUCTURES Excellence Cluster). The authors acknowledge the usage of the AI-clusters Tom and Jerry funded by the Field of Focus 2 of Heidelberg University. The authors are grateful to Paola Ziero for the design of the project logo and their creative support.

References

The references listed include Aarseth et al. 1974, Alvey et al. 2024, Binney & Tremaine 2008, Bovy 2014, Bowden et al. 2015, Bradbury et al. 2018, Buck et al. 2025a and b, Chen et al. 2025, Cranmer et al. 2020, Dax et al. 2025a and b, Durkan et al. 2019, Fardal et al. 2015, Geffner et al. 2022, Gunes et al.

Improvements for AI systems

Amortized Inference Pipeline for Galactic Structure: The improved system can jointly infer host galaxy gravitational potential parameters (like NFW halo mass and scale radius) and progenitor properties (like Plummer mass and scale radius) from single stellar stream phase-space data without requiring sequential training, which is a significant computational gain.
Joint Constraint on Progenitor Physics: The system can capture the intrinsic coupling between tidal stripping dynamics and the underlying potential by jointly constraining parameters that govern both the host galaxy's gravity and the progenitor's internal structure, moving beyond traditional methods like orbit-fitting.
Robust Parameter Inference via Flow Matching: By leveraging conditional Flow Matching to learn a vector field vphi, the system enables efficient, amortized inference directly from stream phase-space data, providing a scalable pathway for dynamical inference that is flexible and avoids the computational expense of gold standard MCMC methods.
Enhanced Model Validation through Simulation-Based Checks: The improved system includes rigorous validation using Predicted Percentile plots for calibration, True-Predicted plots for accuracy, and Posterior Predictive Checks to ensure selfconsistency when forward modeling the inferred parameters against mock observations.
Capturing Complex Correlations: The model can capture intricate dependencies between various parameters, such as correlations between host potential parameters (like MNFW and rNFW) and progenitor mass/scale length parameters, reflecting physical expectations derived from the simulation.

Abstract

Stellar streams offer one of the most sensitive probes of the Milky Way`s gravitational potential, as their phase-space morphology encodes both the tidal field of the host galaxy and the internal structure of their progenitors. In this work, we introduce a framework that leverages Flow Matching and Simulation-Based Inference (SBI) to jointly infer the parameters of the GD-1 progenitor and the global properties of the Milky Way potential. Our aim is to move beyond traditional techniques (e.g. orbit-fitting and action-angle methods) by constructing a fully Bayesian, likelihood-free posterior over both host-galaxy parameters and progenitor properties, thereby capturing the intrinsic coupling between tidal stripping dynamics and the underlying potential. To achieve this, we generate a large suite of mock GD-1-like streams using our differentiable N-body code Odisseo, sampling self-consistent initial conditions from a Plummer sphere and evolving them in a flexible Milky Way potential model. We then apply conditional Flow Matching to learn the vector field that transports a base Gaussian distribution into the posterior, enabling efficient, amortized inference directly from stream phase-space data. We demonstrate that our method successfully recovers the true parameters of a fiducial GD-1 simulation, producing well-calibrated posteriors and accurately reproducing parameter degeneracies arising from progenitor-host interactions. Flow Matching provides a powerful, flexible framework for Galactic Archaeology. Our approach enables joint inference on progenitor and Galactic parameters, capturing complex dependencies that are difficult to model with classical likelihood-based methods.